For example, AM  has a  rule which says  that
whenever  the  "looking  for  extremals"  heuristic  is used  ("in  a
situation f:A→B,  given  an  extremal b⊗6ε⊗*B,  consider  the  subset
A⊗6'=f⊗*⊗A-1⊗*(b)" ), then  AM should construct a  heuristic like the
following and  tack it onto the new concept A': "A' is a useful space
upon which to  test any conjectures involving  f or f⊗A-1⊗*". In  the
case  of f="number-of-factors-of", A  and B  = Natural  numbers, b=2,
this rule  says  to  construct the  heuristic  "Primes  are  valuable
natural   numbers   upon  which   to   test   conjectures   involving
multiplication and  factorization". This heuristic is then associated
with the the newly-created Primes concept. So the only heuristics put
in  by  hand pertain  to  initially-supplied  concepts. All  the  new
heuristics should be producable from those by AM. In reality, this is
not always true; vast improvements in AM's performance are  seen if a
human  user occasionally supplies  AM with an  unperceived heuristic.
Note again the synergetic character of the user--system dialogue.